Skip to content

Research program

Problems money alone
does not solve.

For companies and people working with a fixed budget, we study what it takes to actually run AI.

The problem

  • Hardware

    Most AI models run on datacenter GPUs.

    The machines companies and people actually have sit far below that.

  • Numbers

    Published benchmarks always look good.

    There is rarely a way to check whether the numbers came from the same conditions.

  • Language

    Non-English languages get less attention.

    Most tooling is built around English. Korean spends more tokens on the same sentence, and India's 22 languages have no numbers to compare.

  • Rights

    Labeled open source or open weights does not mean you can use it as-is.

    Whether you can actually run it depends on the license and the provenance.

Approach

We go where the budget does not point

Most AI progress is bought: more parameters, more compute, more data. The frontier moves in whatever direction the largest budget points. Problems in that direction get solved eventually if you wait. We work only on the ones that waiting does not solve.

  1. Finish on the device

    No GPU. No network. Run end to end on hardware people already own.

  2. Keep every result

    One number per test set. Wins, ties, and losses all stay public.

  3. Start with language

    Measure token cost and performance again in Korean and India's 22 languages. Recheck what worked in English.

Licensing

What we can promise

We bind our own work. We do not warranty the upstream.

We bind

  • Apache-2.0 or MIT bases only
  • No training on commercial API outputs
  • Releases under Apache-2.0, as-is

We do not vouch

  • Upstream data we never saw
  • Anything beyond license + model card + stated provenance

Inside a company

Production as the agenda

Constraints from shipping products set the agenda. Artifacts go back into the products.

  1. Products

    Real users, real devices

  2. Constraints

    What spending cannot buy past

  3. Lab artifacts

    Open methods under pressure

Never leaves: customer data, internal costs, pricing, or terms - paper, repo, or benchmark.

Leadership

Janghoon Lee

Head of Redrob Labs, Redrob CTO

janghoonlee.comLinkedInGitHub

Led by one person, but built with 50+ Redrob engineers across Seoul, India, and the US. This is where the foundations of the tools in the Redrob platform get built.

Our best results kept coming from constraints. Trying to buy past them never helped.

Updated August 7, 2026